Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Point-MaDi: Masked Autoencoding with Diffusion for Point Cloud Pre-training
Authors: Xiaoyang Xiao, Runzhao Yao, Zhiqiang Tian, Shaoyi Du
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on Scan Object NN, Model Net40, Shape Net Part, S3DIS, and Scan Net demonstrate that Point-Ma Di achieves superior performance across downstream tasks, surpassing Point-MAE by 5.50% on OBJ-BG, 5.17% on OBJ-ONLY, and 4.34% on PB-T50RS for 3D object classification on the Scan Object NN dataset. |
| Researcher Affiliation | Academia | Xiaoyang Xiao1, Runzhao Yao1, Zhiqiang Tian2, Shaoyi Du1 1 State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi an Jiaotong University 2 School of Software Engineering, Xi an Jiaotong University EMAIL, EMAIL EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methods in detailed textual explanations and diagrams (Figure 1, Figure 2) but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | Codes are available at https://github.com/Yang Parky/Point-Ma Di. |
| Open Datasets | Yes | Extensive experiments on Scan Object NN, Model Net40, Shape Net Part, S3DIS, and Scan Net demonstrate that Point-Ma Di achieves superior performance across downstream tasks, surpassing Point-MAE by 5.50% on OBJ-BG, 5.17% on OBJ-ONLY, and 4.34% on PB-T50RS for 3D object classification on the Scan Object NN dataset. |
| Dataset Splits | Yes | Model Net40 [56] includes 12,311 clean 3D CAD objects with 40 different categories; these objects are split into 9,843 samples in the official training set and 2,468 in the test set. Shape Net Part [60] dataset contains 14,007 and 2,874 samples with 16 object categories and 50 semantic parts for training and validation. S3DIS [2] consists of 3D scan data from 271 rooms across 6 different indoor spaces, which are annotated into 13 classes. We evaluate our model on Area 5, while the other areas are used for fine-tuning our model. |
| Hardware Specification | Yes | All experiments are conducted on a single Ge Force RTX 3090. |
| Software Dependencies | No | The paper mentions using "Py Torch optimizer implementation" but does not specify a version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | We sample each input 1,024 points and divide them into 64 groups, each containing 32 points. We apply scale and translation operations, followed by rotation for data augmentation. The model is pre-trained with a batch size of 128 for 300 total epochs. Following Point-BERT [62], we set the hidden dimension of each encoder block to 384, the number of heads to 6, and the FFN expansion ratio to 4. The depth of the transformer decoder is set to 4. During pre-training, we adopt the Adam W optimizer with a weight decay of 0.05 and an initial learning rate of 5 10 4 with the cosine decay. All experiments are conducted on a single Ge Force RTX 3090. To ensure a fair comparison, we employed identical experimental settings to the default fine-tuning. More details are provided in Tab. 7. |